One important correction first: the statement that Anthropic is “targeting 5 GW by the end of 2026” is too broad. Anthropic has secured up to 5 GW of new AWS capacity, but its own announcement says nearly 1 GW of Trainium2/Trainium3 capacity is expected to come online by the end of 2026; the broader 5-GW commitment is a longer-term infrastructure expansion.
1. What is the “AI chip race”?
Think about AI development as a chain:
AI models → computing → AI chips → servers → networking → data centers → electricity + cooling
Companies such as Anthropic, OpenAI and Google are competing to build increasingly powerful AI models.
But there is a problem:
The better the AI model becomes, the more computing power it generally needs.
That means the companies cannot simply write better software. They also need massive amounts of hardware.
And this is why the chip race has become one of the biggest stories in technology and markets.
2. Why NVIDIA is no longer the entire story
For years, the simplified AI-chip story was:
AI → NVIDIA GPUs
NVIDIA remains enormously important, but AI companies increasingly want alternatives and customization.
Why?
Because relying entirely on one type of accelerator can mean:
- high costs
- limited supply
- power inefficiency for particular workloads
- dependence on one supplier
- less control over hardware/software optimization
So the industry is moving toward a much more complicated ecosystem.
The emerging landscape
| Company | Strategy |
|---|---|
| NVIDIA | General-purpose AI accelerators + networking |
| AMD | Alternative AI accelerators |
| Custom TPU accelerators | |
| Amazon/AWS | Trainium custom AI chips |
| Broadcom | Custom AI accelerator design + networking |
| OpenAI | Developing its own AI accelerators |
| Anthropic | Using multiple chip platforms |
| Microsoft | AI infrastructure + custom silicon/cloud ecosystem |
The important trend is:
AI companies are becoming increasingly involved in designing the hardware they run on.
3. Anthropic: the “5 GW” story
This is one of the most important parts.
Anthropic has dramatically increased its computing ambitions.
In April 2026, Anthropic and Amazon announced an agreement for up to 5 gigawatts of new compute capacity for training and deploying Claude.
Anthropic says the commitment involves more than $100 billion over ten years in AWS technologies. (Anthropic)
But here’s the important distinction:
5 GW ≠ 5 GW online by December 2026.
Anthropic said that nearly 1 GW of Trainium2 and Trainium3 capacity is expected to come online by the end of 2026. (Anthropic)
Anthropic is also diversifying its hardware.
It currently uses:
AWS Trainium + Google TPUs + NVIDIA GPUs
rather than relying exclusively on one chip platform. (Anthropic)
4. Why 5 GW is such a huge number
Most people hear 5 gigawatts and don’t immediately understand what it means.
Think of it as an enormous amount of continuous electrical capacity dedicated to computing infrastructure.
AI data centers aren’t just warehouses full of computers.
They require:
AI chips
↓
Servers
↓
Networking
↓
Cooling
↓
Power infrastructure
↓
Data-center buildings
So when Anthropic says it needs gigawatts of compute capacity, the implications extend far beyond semiconductor companies.
It creates demand for:
- GPUs
- custom accelerators
- HBM memory
- networking chips
- optical networking
- servers
- transformers
- power systems
- cooling equipment
- data centers
- electricity generation
- cloud infrastructure
That’s why the AI boom is increasingly being described as an infrastructure buildout.
5. Anthropic is also betting on Google TPUs
Anthropic isn’t putting all its chips in Amazon’s basket.
In April 2026, Anthropic announced a separate agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity, expected to begin coming online in 2027. (Anthropic)
This is strategically important.
Google designs its own AI accelerators called TPUs — Tensor Processing Units.
So instead of:
“Everybody buys NVIDIA GPUs.”
the industry is moving toward:
NVIDIA + Google TPU + Amazon Trainium + AMD + custom accelerators
That is much more competitive.
6. And then there is OpenAI
OpenAI is taking a similar path.
In October 2025, OpenAI and Broadcom announced a collaboration to deploy 10 gigawatts of OpenAI-designed AI accelerators.
The systems are targeted to begin deployment in the second half of 2026, with deployment continuing through 2029. (Broadcom Inc.)
This is a massive strategic shift.
OpenAI isn’t simply saying:
“Give us more GPUs.”
It is effectively saying:
“We want to help design the computing infrastructure that powers our models.”
7. Why Broadcom is suddenly extremely important
This is where your earlier AI infrastructure story connects directly.
Broadcom isn’t primarily known by consumers like NVIDIA is.
But it sits underneath a huge part of the AI infrastructure ecosystem.
Broadcom works on:
- custom AI accelerators
- networking
- Ethernet
- connectivity
- data-center infrastructure
The OpenAI partnership involves both AI accelerators and networking systems. (Broadcom Inc.)
So Broadcom represents an important trend:
Hyperscalers and AI labs increasingly want custom silicon.
Instead of buying a completely standardized processor, a company can work with a chip designer to create hardware optimized for its particular workload.
8. What does “custom AI chip” actually mean?
Imagine two cars.
NVIDIA GPU
Like a high-performance sports car that can perform many different jobs.
It is extremely powerful and flexible.
Custom AI accelerator
More like a specialized racing car designed for one particular type of track.
It may not be as flexible, but it can potentially deliver:
better efficiency + lower cost + optimized performance
for a specific workload.
This becomes extremely important when you’re operating millions of AI queries every day.
Even a small efficiency improvement can become worth billions of dollars at massive scale.
9. OpenAI’s custom chip is already more than just a future idea
This is another place where the original wording needs updating.
OpenAI unveiled its first custom AI chip, designed with Broadcom, in June 2026.
The chip, called Jalapeño, is designed specifically for AI inference—the stage where a trained model actually processes a user’s request and produces an answer.
So OpenAI’s hardware strategy has moved from:
“We might make chips someday.”
to:
“We are actually designing and deploying custom silicon.”
That is a major development.
10. Where does Google fit?
Google is one of the most vertically integrated players in AI.
It has:
AI models → Gemini
AI chips → TPUs
Cloud → Google Cloud
Data centers → Google infrastructure
Applications → Search, Workspace, Android, etc.
That gives Google something many AI startups don’t have:
Control over almost the entire AI stack.
And Anthropic’s growing use of Google TPUs makes Google’s chip business even more strategically important. Anthropic previously announced plans to expand to as many as one million Google TPUs, with more than 1 GW expected to come online in 2026. (Anthropic)
11. This creates a fascinating three-way competition
You can simplify the current situation like this:
🟢 Anthropic
Claude
↓
Needs enormous compute
↓
AWS Trainium + Google TPU + NVIDIA
↓
Diversified infrastructure
🔵 OpenAI
ChatGPT
↓
Needs enormous compute
↓
NVIDIA + AMD + custom OpenAI/Broadcom accelerators
↓
Increasing hardware independence
Gemini
↓
Own models
↓
Own TPUs
↓
Own cloud
↓
Own data centers
↓
Extremely vertically integrated AI ecosystem
12. Why NVIDIA should still not be ignored
This doesn’t mean NVIDIA is suddenly irrelevant.
Quite the opposite.
Even companies developing custom chips continue to use NVIDIA GPUs.
Anthropic explicitly says it uses AWS Trainium, Google TPUs and NVIDIA GPUs. (Anthropic)
And AWS and NVIDIA announced in August 2026 that AWS plans to deploy 2 million additional NVIDIA GPUs across its global infrastructure. (US Press Center)
So the story isn’t:
Custom chips will kill NVIDIA.
It’s more accurately:
The AI accelerator market is becoming more diversified.
NVIDIA can remain dominant while Google TPUs, AWS Trainium and custom ASICs take specific workloads.
13. The real battle: cost per AI query
This is perhaps the most important concept for understanding the next stage.
Training a model is expensive.
But once millions or billions of people use that model, inference costs become enormous.
Imagine:
1 million users → manageable
100 million users → huge
1 billion users → gigantic infrastructure requirement
At that scale, saving even a few cents per workload can become extremely valuable.
That’s why custom inference chips are so interesting.
The future AI competition isn’t only:
Who has the smartest model?
It increasingly becomes:
Who can run the smartest model at the lowest cost and highest reliability?
14. And that’s where energy becomes a massive issue
Here’s the connection many people miss.
More AI models →
more computation →
more chips →
more servers →
more data centers →
more electricity →
more cooling.
Therefore, the AI boom is also becoming an energy and infrastructure story.
This creates opportunities and challenges for companies involved in:
- electricity generation
- grid infrastructure
- data-center construction
- cooling
- power management
- networking
- semiconductor manufacturing
This is why your previous TSMC and AI infrastructure topics fit perfectly into the same overall narrative.
15. TSMC is sitting underneath much of this
Now connect everything:
Anthropic
OpenAI
Google
NVIDIA
AMD
Broadcom
Amazon
↓
Need advanced chips
↓
Advanced semiconductor manufacturing
↓
TSMC and other foundries
That’s why TSMC’s revenue data is an important market signal.
If AI companies continue spending aggressively on compute, the effects eventually show up throughout the semiconductor supply chain.
16. The bigger investment story
The market is gradually moving from:
Phase 1
“Who has the best AI model?”
to:
Phase 2
“Who has enough compute?”
to:
Phase 3
“Who can operate that compute most efficiently?”
And Phase 3 is where custom silicon, networking, memory, power and cooling become increasingly important.
17. What investors should watch
For your blog, I’d track these five areas:
① AI capital expenditure
Are Microsoft, Amazon, Google and Meta continuing to increase infrastructure spending?
② Custom chip adoption
Are OpenAI, Anthropic and hyperscalers actually shifting more workloads toward their own/custom accelerators?
③ NVIDIA’s position
Does NVIDIA continue to maintain performance and ecosystem advantages?
④ Semiconductor demand
Watch:
TSMC → Broadcom → NVIDIA → AMD → Micron → semiconductor equipment companies
⑤ Power availability
This could become one of the biggest bottlenecks.
Having money to build an AI data center doesn’t help if you can’t get enough electricity to operate it.
🎯 The simplest way to explain the entire story
The AI race is no longer just about building smarter models. It’s becoming a race to build enough computing capacity—and to make that computing cheaper, faster and more energy-efficient.
Anthropic is expanding across AWS Trainium, Google TPUs and NVIDIA GPUs, including an agreement for up to 5 GW of AWS capacity, while OpenAI is moving toward 10 GW of custom accelerators with Broadcom; meanwhile, Google continues developing its own TPU ecosystem. (Anthropic)
“The next AI winner may not simply be the company with the smartest model—it may be the company that can run intelligence at the lowest cost and at the largest scale.”
That is the real story behind the Anthropic–OpenAI–Google chip race.
